Daniele Montanino

Univ. del Salento & INFN

Zone refining
 Zone refining
• Zone refining (or zone melting) was invented in the Early
50’s by J.D. Bernal and developed by W.G. Pfann in Bells
Laboratories to prepare high-purity materials (mainly
semiconductors)
• A narrow region of a crystal is melted, and this molten
zone is moved through the crystal. The molten region
melts impure solid at its forward edge and leaves a wake
of purer material solidified behind it as it moves through
the ingot.

The Journal of The Minerals, Metals & Materials Society 4, 747 (1952)
 Oven
Quartz ampoule

Impurities tend to
remain preferably (but
not always!) in the liquid
Molten zone
zone and thus are
dragged from tip to tail.

The procedure can be

repeated many times to
increase the purity of
the ingot.

After cutting the tail we

remain with a purified
ingot.
After solidification in the interface between
liquid and solid a fraction k (coefficient of
segregation) of impurities contained in the
liquid remain in the solid

𝐶𝑆 = 𝑘𝐶𝐿

where Cs (CL) is the concentration of pollutant

per volume unit in the solid (liquid).

If k<1 (k>1) the pollutant remains preferably in

the molten (frozen) zone.
Theory of zone refining
w
dx dx

x x+w
L
𝑤𝑆𝑑𝐶𝐿 = Variation of solute in the molten zone

−𝐶𝑠 𝑥 𝑆𝑑𝑥 + 𝐶𝑠 𝑥 + 𝑤 𝑆𝑑𝑥

Solute entering in the freezing zone Solute entering in the molten zone
Remembering that at point x the variation of the solute entering in the freezing zone is
𝑑𝐶𝑆 (𝑥) = 𝑘𝑑𝐶𝐿 (𝑥)
and thus
𝑑𝐶𝑆,after (𝑥) 𝑘
= 𝐶𝑆,before 𝑥 + 𝑤 − 𝐶𝑆,after (𝑥)
𝑑𝑥 𝑤

Previous equation is valid only if the pollutant diffuses uniformly in the molten zone.
The effective segregation coefficient in the process is given by
𝑘
𝑘eff =
𝑉𝛿
𝑘 + 1 − 𝑘 exp −
𝐷

with V=moving rate of the oven, D=diffusion coefficient, =thickness of the diffusion
boundary layer (fraction of mm). Diffusion coefficient in molten salts is usually
O(10-5 cm2/s). If V<D/ ~ mm/h we have keff➝k. Convection due to temperature
gradient and Marangoni convection can contribute to diffusion.
Tail zone: when the oven reaches and overcomes the end of ingot, the equation is
different
L-x

x
L
𝑆𝑑 (𝐿 − 𝑥)𝐶𝐿 = −𝐶𝑠 𝑥 𝑆𝑑𝑥
Variation of solute in the molten zone = solute entering in the freezing zone

𝑑𝐶𝑆 (𝑥) 1 − 𝑘 𝐿−𝑥 𝑘−1

= 𝐶𝑆 (𝑥) 𝐶𝑆 (𝑥)= 𝐶𝑆 (𝐿 − 𝑤)
𝑑𝑥 𝐿−𝑥 𝑤
Starting condition: in the tip enters a fraction k of the average concentration of the
molten zone
w

𝑤
𝑘
𝐶𝑆,after 0 = 𝑘𝐶𝑆,before = න 𝐶𝑆,b𝑒fore 𝑥 𝑑𝑥
𝑤
0
Recap: suppose to repeat the procedure n times. The equations to solve are

𝑘
𝐶𝑛′ 𝑥 = 𝐶𝑛−1 𝑥 + 𝑤 − 𝐶𝑛 𝑥 if 𝑥 ≤𝐿−𝑤
𝑤
𝑘−1
𝐿−𝑥
𝐶𝑛 𝑥 = 𝐶𝑛 𝐿 − 𝑤 if 𝐿 − 𝑤 < 𝑥 ≤ 𝐿
𝑤

𝑘 𝑤
𝐶𝑛 0 = න 𝐶𝑛−1 𝑥 𝑑𝑥
𝑤 0

We suppose that initially the ingot has a constant concentration C0(x)=C0.

The functions fn(x)=Cn(x)/C0 follow the previous equations but with f0=1.
For f1(x) we have a simple analytic solution
𝑘𝑥
−
𝑓1 𝑥 = 1 − 1 − 𝑘 𝑒 𝑤 if 𝑥 ≤𝐿−𝑤
Anyway, for n>1 the equations must be solved through a numerical iterative procedure

Yes
𝑛 = 0, 𝑓𝑛 𝑥 =1 𝑛 →𝑛+1 𝑛 = 𝑛max ? END

No

𝑘
𝑓𝑛 0 = ෍ 𝑓𝑛−1 𝑥𝑖 Δ𝑥
𝑤
𝑥𝑖 ≤𝑤
𝑘
𝑓𝑛 𝑥𝑖+1 = 𝑓𝑛 𝑥𝑖 + 𝑓 𝑥 + 𝑤 − 𝑓𝑛 𝑥𝑖 Δ𝑥 for 𝑥𝑖 ≤ 𝐿 − 𝑤
𝑤 𝑛−1 𝑖
𝐿 − 𝑥𝑖 𝑘−1
𝑓𝑛 𝑥𝑖 = 𝑓𝑛 𝐿 − 𝑤 if 𝐿 − 𝑤 < 𝑥𝑖 ≤ 𝐿
𝑤
The longer is the ampoule the better is!
Limit (“ultimate”) solution: repeating the procedure to infinity is useless. In fact,
for increasing n a limiting solution is reached. For n➝∞ the equation become
𝑘
𝑓∞′ 𝑥 = 𝑓∞ 𝑥 + 𝑤 − 𝑓∞ 𝑥
𝑤
this equation can be solved by an exponential

𝑓∞ 𝑥 = 𝐴𝑒 𝐵𝑥/𝑤

B is the solution of the transcendent equation 𝐵 = 𝑘 𝑒 𝐵 − 1 and A can be found

imposing the conservation of the solute

1 𝐿 𝐵𝐿/𝑤
න 𝑓∞ 𝑥 𝑑𝑥 = 1 ⟹ 𝐴 = 𝐵𝐿/𝑤
𝐿 0 𝑒 −1

However, this is only an approximate solution since for x>L-w the solution is a
power-law and not exponential!
x x+w

f∞(x+w)
not an exponential
Geometry of the ampoule: the ampoule has a cone-shaped tip

we must take in account the varying cross section of the ampoule.

w

A(x) S

x
Xc
If V(x) is the volume of the molten zone and A(x) the cross section of the ampoule at the
left side, the equation become
𝑑 𝑉(𝑥)𝐶𝐿 = 𝐶𝑠 𝑥 + 𝑤 𝑆𝑑𝑥 − 𝐶𝑠 𝑥 𝐴(𝑥)𝑑𝑥
After a little algebra we obtain the corrected equation

𝑆 𝐴 𝑥
𝐶𝑛′ 𝑥 = 𝑘𝐶𝑛−1 𝑥 + 𝑤 − 1 − 1 − 𝑘 𝐶𝑛 𝑥
𝑉(𝑥) 𝑆

2 1 𝑥2
𝑤 − 𝑋𝑐 + 𝑥 1 − if 𝑥 ≤ 𝑋𝑐
𝑉(𝑥) = 𝑆 ∙ ൞ 3 3 𝑋𝑐2
𝑤 𝑥 > 𝑋𝑐

𝑥2
if 𝑥 ≤ 𝑋𝑐
𝐴 𝑥 = 𝑆 ∙ ൞𝑋𝑐2
1 𝑥 > 𝑋𝑐
Notice that previous equation can be generalized to any geometry of the ampoule. The
most general equation is

𝐶𝑛′ 𝑥 𝑉 𝑥 = 𝑘 𝐶𝑛−1 𝑥 + 𝑤 𝐴 𝑥 + 𝑤 − 𝐶𝑛 𝑥 𝐴 𝑥 − 𝐶𝑛 𝑥 𝐴 𝑥 + 𝑤 − 𝐴 𝑥

where A(x) is the cross section of the ampoule at the point x and V(x) the volume of the
molten zone:
𝑥+𝑤
𝑉 𝑥 =න 𝐴 𝑥 𝑑𝑥
𝑥

and the starting condition

𝑤
1
𝐶𝑛 0 = න 𝐶𝑛−1 𝑥 𝐴 𝑥 𝑑𝑥
𝑉(0) 0

This equation includes also the tail, in which V(x) =S∙(L-x) and A(x+w)=0 if x>L-w.
Cutting the edge: after refining the tail of the ingot is cut away and the rest of the
ingot is shuttered and recast in the final crystal growth (normal freezing). The
average contaminant concentration is a function of the cut

𝐶0 𝑋
ҧ
𝐶(𝑋) = න 𝑓𝑛 𝑥 𝑑𝑥
𝑋 0

X
 𝐶0 𝑋
ҧ
𝐶(𝑋) = න 𝑓 𝑥 𝑑𝑥
𝑋 0 𝑛
 𝐶0 𝑋
ҧ
𝐶(𝑋) = න 𝑓 𝑥 𝑑𝑥
𝑋 0 𝑛
Application to SABRE: ultrapure NaI crystals are necessary to compete with DAMA/LIBRA
Radioclean NaI powder Astrograde by Sigma Aldrich
(now Merck) is refined at Mellen company (Concord, NH)

The purified (but raw) crystal is moved to

Radiation Monitoring Device (Boston)

Final crystal is For each run 5 samples from ingot are taken at the position in figure
grown in vertical and shipped at LSC (Canfranc) and Seastar for analysis
Bridgman method
in fused silica 1’’
vessels (normal
freezing)
Zone #1 #2 #3 #4 #5
 ZR Run 1 (26 passes, 1.5 inch/h, ampoule with Carbon coating, powder: 76650)
Sample 39K 65Cu 85Rb 133 Cs 138 Ba 88Sr 208 Pb

[ppb] [ppb] [ppb] [ppb] [ppb] [ppb] [ppb]

LSC Seast LSC Seast LSC Seast LSC Seast LSC Seast Seast LNGS LSC LSC Seast LNGS
powder 14.4 6.7 <5 <4 <0.8 <0.4 11 19 <0.6 0.18 1 2 1.6
Zone #1 <4 <4 <5 <3 <0.8 <0.4 <0.3 <2 <0.5 <0.3 367 350 5.1 3.6
Zone #2 <4 <4 <5 <3 <0.8 <0.4 <0.3 <2 0.8 <0.3 287 180 93 2.6 2.9 2.5
Zone #3 6.7 <4 <5 <3 <0.8 <0.4 0.4 <2 0.8 <0.3 86 <1 1.8
Zone #4 40 6.3 <5 <3 <0.8 <0.4 0.4 <2 3.8 <0.3 41 20 <1 1.2 1.6
Zone #5 540 275 234 150 1.3 <0.4 447 570 10 3.7 10 4 4.7 1.9 0.95 0.9

The two main contaminants for the experiment are 40K and 210Pb
Statistical analysis: in order to find the segregation constants k for the various
contaminant previous data must be fitted with the ZR model discussed previously.

The fitting procedure procedure must face three problems

➢ The first zones present only upper limits due to the sensitivity of the procedure
of measurement
➢ The measures are carried out on grains taken randomly in each zone after
crunching. For this reason, the measures have a uniform uncertainty on the x
position
➢ The error on the measures is unknown. We assume conservatively a 20%
gaussian error on the measures (dominated by systematics) while for upper
limits we assume a uniform probability for all values smaller that the limit

For these reasons, a maximum likelihood procedure is more appropriate for fitting
rather than a 2.
Fitting procedure: we can write the likelihood function as

𝐿 𝑘, 𝐶0 = 𝑃𝑟(𝐶0 ) × ෑ 𝒫𝑧 𝑘, 𝐶0
𝑧=1

where 𝒫𝑧 are the probabilities in each zone and Pr is a prior. For zones in which we
have a measure Czexp
𝛿
𝑥𝑧 + 𝑧
2 𝑒𝑥𝑝 2
𝐶𝑧 − 𝐶0 𝑓𝑛 (𝑥)
𝒫𝑧 𝑘, 𝐶0 ∝ න exp − 𝑑𝑥
2𝜎𝑧2
𝛿
𝑥𝑧 − 𝑧
2

with xz and z central value and width of the z-th bin, z estimated error on the
measure (we fix z=0.2×Czexp) and n number of passes. We thus consider all the
values x ∈[xz-½z,xz+½z] equiprobable.
For the zones in which we have only an upper bound Czbound we have:

a) If C0fn(x) > Czbound ∀ x ∈[xz-½z,xz+½z] we put 𝒫𝑧 =0 since the theoretical values

cannot be higher than the upper value
b) Conversely, if C0fn(x) < Czbound ∀ x ∈[xz-½ z,xz+½z] we have 𝒫𝑧 =1 since each value
of x in the interval is equiprobable
c) If ∃ ∈[xz-½z,xz+½z] with C0fn() = Czbound we can consider only the fraction of bin in
which the curve is below the bound. E.g., if fn(x) is increasing 𝒫𝑧=[ -(xz-½z)]/z

C0fn(x)
C0fn(x)
𝒫𝑧 =0 𝒫𝑧 =1
Czbound
xz-½ z xz+½z xz-½z xz+½z xz- ½z xz+ ½z

C0fn(x) 𝒫𝑧 = [ -(xz-½z)]/z
a) b) c)

xz xz xz 
Finally, in order to avoid that the fitted value of C0 deviate too much from the value
measured in the powder we multiply by a prior function

2
𝐶0 − 𝐶𝑝𝑜𝑤𝑑𝑒𝑟
𝑃𝑟 𝐶0 = exp − 2
2𝜎𝑝𝑜𝑤𝑑𝑒𝑟

(we assume powder=0.2×Cpowder).

We define the likelihood ratio function

𝐿 𝑘, 𝐶0
𝜆 𝑘, 𝐶0 = −2log
෠ 𝐶መ0
𝐿 𝑘,

were 𝑘෠ and Ĉ0 are the value that maximize the likelihood (best fit values). According to
the Wilk’s theorem this function behaves approximatively like a 2.
 0.2

1 Powder
1
90%

99%

k=0.58±0.05 @ 1
 0.2

1 Powder

k=0.41±0.08 @ 1
Excluding last bin doesn’t change too much the results…
Excluding last bin doesn’t change too much the results…
1 Powder

k=1.15±0.03 @ 1
 4

1 Powder

k=0.68±0.02 @ 1
Comparison with literature: we tried to reproduce the analysis done by Suerfu,
Calaprice and Souza [Phys. Rev. Appl. 16, 014060, (2021)]. We can reproduce exactly
their Fig. 4...

Suerfu et al.
 Table I
…however, we cannot reproduce their results in Fig. 5!

w/L=0.057?
0.08?
Anyway, using their data we found more or less the same best fit. We assume again a
20% error on measure (in Suerfu et al. the error is not specified). We also assumed the
“a posteriori” error quoted by Suerfu et al. on C0 (5.3% error).
 Fig. 6 by Suerfu et al. vs our analysis.
Notice that the scales are different

1 Powder
k=0.571±0.006 @ 1 k=0.56±0.02@@1
k=0.56±0.03 1

We cannot reproduce the analysis by Suerfu et al. due

to lack of information on the analysis procedure and
some incongruences in their figures. The errors on k
and C0 seem to be underestimated.
Thank you for your attention
This work has been performed in collaboration with

• Gabriella Cataldi
• Samina Zuhra
• Sanagul Khattak
• Chemseddine Ananna
• Maria Luisa De Giorgi